Automatic driving model and automatic driving method

Abstract

The invention discloses an automatic driving model and an automatic driving method, and relates to the technical field of automatic driving. The automatic driving model comprises a feature extraction and coding module, an attention anti-coding module and a graph neural network module. The feature extraction and coding module is connected with the attention anti-coding module, the feature extraction and coding module is used for performing feature extraction and coding on the driving information of the vehicle, and the attention anti-coding module is used for decoding feature coding information output by the feature extraction and coding module; the attention anti-coding module is connected with the graph neural network module, and the graph neural network module is used for predicting potential motion behaviors of the vehicle according to a decoding result output by the attention anti-coding module; therefore, forward propagation is only carried out on the GPU so as to complete output of various task requirements in automatic driving, the problem that data transmission needs to be carried out between the CPU and the GPU is solved, the data processing efficiency of an automatic driving model is improved, and the data transmission cost is reduced.

Description

Technical Field

[0001] The present application relates to the field of autonomous driving technology, and in particular to an autonomous driving model and an autonomous driving method. Background Art

[0002] Autonomous driving technology usually includes technologies such as high-precision maps, environmental perception, path planning and path tracking control, which enable the vehicle to automatically drive to the destination without human control; among them, the autonomous driving model is the key to the autonomous driving vehicle in the autonomous driving process, and can provide autonomous driving vehicles with driver-like decision-making, operation or control in actual driving scenarios.

[0003] In view of this, the safe driving of autonomous vehicles depends on the accuracy and stability of the autonomous driving model; in particular, the current autonomous driving model usually relies on the collaborative processing of the central processing unit (CPU) and the graphics processing unit (GPU), and can already complete the output of many task requirements in autonomous driving.

[0004] However, the use of the above-mentioned autonomous driving model requires data transmission between the CPU and the GPU to achieve the output of the task requirements in autonomous driving, which takes a lot of time, thereby reducing the data processing efficiency of the autonomous driving model and increasing the data transmission cost.

[0005] Therefore, using the above method, the data processing efficiency of the autonomous driving model is low and the cost of data transmission is high. Summary of the invention

[0006] The embodiments of the present application provide an autonomous driving model and an autonomous driving method to solve the problem of data transmission between the CPU and the GPU during the autonomous driving process, thereby improving the data processing efficiency of the autonomous driving model and reducing the cost of data transmission.

[0007] In a first aspect, an embodiment of the present application provides an autonomous driving model, which is applied to a graphics processor (GPU), and the autonomous driving model includes: a feature extraction and encoding module, an attention de-encoding module, and a graph neural network module;

[0008] The feature extraction and encoding module is connected to the attention de-encoding module; wherein the feature extraction and encoding module is used to extract and encode the driving information of the vehicle, and the attention de-encoding is used to decode the feature encoding information output by the feature extraction and encoding module;

[0009] The attention de-coding module is connected to the graph neural network module; wherein, the graph neural network module is used to predict the potential movement behavior of the vehicle based on the decoding result output by the attention de-coding module.

[0010] In the above embodiment, the feature extraction and encoding module is connected to the attention de-encoding module, and the attention de-encoding module is connected to the graph neural network module; in this way, feature extraction and encoding of the vehicle's driving information can be achieved, and the feature encoding information can be decoded, so that according to the decoding results, the output of many task requirements in autonomous driving can be completed; and the autonomous driving model is applied to the GPU, and only forward propagates on the GPU to complete the output of many task requirements in autonomous driving; in this way, the problem of data transmission between the CPU and the GPU during the autonomous driving process is solved, thereby improving the data processing efficiency of the autonomous driving model and reducing the cost of data transmission.

[0011] In an optional embodiment, the autonomous driving model also includes: a global positioning module connected to the feature extraction and encoding module, and the global positioning module is used to determine the global positioning posture of the vehicle based on the posture information collected by each of the multiple sensors of the vehicle.

[0012] In the above embodiment, the posture information collected by multiple sensors of the vehicle is comprehensively considered to improve the accuracy of the vehicle's global positioning posture, and to a certain extent ensure the accuracy of the output results of many task requirements during the autonomous driving process.

[0013] In an optional embodiment, the autonomous driving model also includes: a three-dimensional 3D vector reconstruction module, which is used to reconstruct a vector map around the vehicle's driving trajectory based on the global positioning posture and the local vector fragments in the decoding results; wherein the local vector fragments include at least: the local positioning posture of the vehicle.

[0014] In the above embodiment, a vector map around the vehicle's driving trajectory is reconstructed based on the global positioning posture and local vector fragments in the decoding results, thereby improving the accuracy of the real-time reconstructed vector map; and the vectorization results are output end-to-end, reducing the data transmission cost.

[0015] In an optional embodiment, the feature extraction and encoding module includes: an image preprocessing module and a temporal attention module; wherein the image preprocessing module is used to perform parameter calibration and feature extraction on the global positioning posture to obtain multiple posture feature maps, and the temporal attention module is used to perform temporal tracking on the multiple posture feature maps.

[0016] In the above embodiment, the image preprocessing module and the temporal attention module are used to preprocess the global positioning posture, and more accurate multiple posture feature maps are obtained. In addition, based on the original spatiotemporal sequence attention mechanism and flexible training reasoning network, the automatic driving model can greatly improve the on-chip reasoning speed without changing the accuracy.

[0017] In an optional embodiment, the temporal attention module is also used to obtain the feature coding information based on multiple pose feature maps after temporal tracking, a set of feature sampling points associated with the vehicle type of the vehicle in a first preset index table, and a set of feature sampling point weights associated with the vehicle type in a second preset index table.

[0018] In the above embodiment, according to the first preset index table and the second preset index table set for the vehicle type, the feature sampling point set and its corresponding feature sampling point weight set are queried, which not only improves the versatility of the autonomous driving model, but also greatly reduces the occupation of on-chip computing resources and computational complexity, thereby improving the efficiency of vehicle-side reasoning.

[0019] In an optional embodiment, the attention de-encoding module includes multiple attention mechanism network layers, each of which is trained according to a corresponding feature sampling point set and a preset loss function.

[0020] In the above embodiment, the attention mechanism network layer is trained according to the feature sampling point set and the preset loss function, which improves the accuracy of the attention de-encoding module and also improves the safety of the autonomous driving model during the autonomous driving process.

[0021] In an optional embodiment, the attention de-encoding module is also used to generate a 3D target and estimate the 3D motion trajectory of the 3D target, and determine the 3D occupied area of ​​the 3D target based on the 3D motion estimation result of the 3D target.

[0022] In the above embodiment, the generation of 3D motion trajectories, 3D motion estimation results and 3D occupied areas of other road participants is realized. In this way, during the autonomous driving process, the status of other road participants is fully considered, further improving the accuracy of subsequent predictions of the vehicle's potential motion behavior.

[0023] In a second aspect, an embodiment of the present application further provides an autonomous driving method of an autonomous driving model, which is applied to a GPU, and the method includes:

[0024] In response to an automatic driving instruction of a target object, obtaining driving information of a vehicle associated with the automatic driving instruction;

[0025] Inputting the driving information of the vehicle into a pre-trained autonomous driving model to obtain an autonomous driving strategy for the vehicle; wherein the autonomous driving strategy includes: potential motion behaviors of the vehicle;

[0026] Based on the autonomous driving strategy, instruct the vehicle to perform autonomous driving.

[0027] The autonomous driving method based on the above-mentioned autonomous driving model responds to the autonomous driving instructions of the target object, obtains the driving information of the vehicle associated with the autonomous driving instructions, and then inputs the driving information of the vehicle into a pre-trained autonomous driving model to obtain an autonomous driving strategy including the potential motion behavior of the vehicle, and then instructs the vehicle to perform autonomous driving based on the obtained autonomous driving strategy; in this way, the output of many task requirements in autonomous driving can be completed only by forward propagation on the GPU, and the problem of data transmission between the CPU and GPU during the autonomous driving process is solved, thereby improving the data processing efficiency of the autonomous driving model and reducing the cost of data transmission.

[0028] In an optional embodiment, inputting the driving information of the vehicle into a pre-trained autonomous driving model to obtain the autonomous driving strategy of the vehicle includes:

[0029] Parsing the driving information to obtain a plurality of position information of the vehicle at the current moment and the environment information of the vehicle at the current moment;

[0030] Based on the multiple pose information, the global positioning pose of the vehicle is determined, and based on the global positioning pose and the environmental information, an automatic driving strategy for the vehicle is generated.

[0031] In an optional embodiment, the process of instructing the vehicle to perform automatic driving based on the automatic driving strategy further includes:

[0032] In response to the target object's instruction to display the driving road condition of the vehicle, the vehicle is instructed to display a 3D motion trajectory and a state control signal within a future set time range.

[0033] In addition, other features and advantages of the present application will be described in the subsequent description, and partly become apparent from the description, or be understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. In the drawings:

[0035] Figure 1 A schematic diagram of the composition structure of an autonomous driving model provided in an embodiment of the present application;

[0036] Figure 2 A schematic diagram of a specific structure of an autonomous driving model provided in an embodiment of the present application;

[0037] Figure 3 A schematic diagram of a framework of a high-precision positioning algorithm provided in an embodiment of the present application;

[0038] Figure 4 A logical schematic diagram of 3D vector map reconstruction provided in an embodiment of the present application;

[0039] Figure 5 A schematic diagram of the composition structure of a feature extraction and encoding module provided in an embodiment of the present application;

[0040] Figure 6 A schematic diagram of algorithm logic for training a temporal attention module provided in an embodiment of the present application;

[0041] Figure 7 A schematic diagram of the algorithm logic of a temporal attention module during vehicle-side reasoning provided in an embodiment of the present application;

[0042] Figure 8 An algorithm logic diagram of an attention de-coding module provided in an embodiment of the present application;

[0043] Fig. 9 An algorithm logic diagram of a graph neural network module provided in an embodiment of the present application;

[0044] Fig.10 A detailed design diagram of a pure spatiotemporal end-to-end autonomous driving model provided in an embodiment of the present application;

[0045] Fig.11 A schematic diagram of the implementation flow of an autonomous driving method for an autonomous driving model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the technical solution of the present application, rather than all of the embodiments. Based on the embodiments recorded in the application documents, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the technical solution of the present application.

[0047] It should be noted that in the description of this application, "multiple" is understood as "at least two". "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent three situations: A exists alone, A and B exist at the same time, and B exists alone. A and B are connected, which can represent two situations: A and B are directly connected and A and B are connected through C. In addition, in the description of this application, words such as "first" and "second" are only used for the purpose of distinguishing descriptions, and cannot be understood as indicating or implying relative importance, nor can they be understood as indicating or implying order.

[0048] In addition, the collection, dissemination, and use of data in the technical solution of this application comply with the requirements of relevant national laws and regulations.

[0049] The following is a brief introduction to the design concept of the embodiment of the present application:

[0050] The safe driving of autonomous vehicles depends on the accuracy and stability of the autonomous driving model; in particular, the current autonomous driving models can usually complete the output of many task requirements in autonomous driving with the help of collaborative processing of CPU and GPU.

[0051] However, the use of the above-mentioned autonomous driving model requires data transmission between the CPU and the GPU to achieve the output of the task requirements in autonomous driving, which takes a lot of time, thereby reducing the data processing efficiency of the autonomous driving model and increasing the data transmission cost.

[0052] Therefore, in order to achieve efficient on-chip computing under full GPU resources, the embodiments of the present application provide an autonomous driving model and an autonomous driving method to solve the problem of data transmission between the CPU and the GPU during the autonomous driving process, thereby improving the data processing efficiency of the autonomous driving model and reducing the cost of data transmission.

[0053] See also Figure 1 As shown, the autonomous driving model includes: feature extraction and encoding module, attention de-encoding module and graph neural network module;

[0054] The feature extraction and encoding module is connected to the attention de-encoding module; wherein the feature extraction and encoding module is used to extract and encode the driving information of the vehicle, and the attention de-encoding is used to decode the feature encoding information output by the feature extraction and encoding module;

[0055] The attention de-coding module is connected to the graph neural network module; wherein the graph neural network module is used to predict the potential motion behavior of the vehicle based on the decoding results output by the attention de-coding module.

[0056] The feature extraction and encoding module is connected to the attention de-encoding module, and the attention de-encoding module is connected to the graph neural network module; in this way, it is possible to extract and encode the vehicle's driving information and decode the feature encoding information, so that according to the decoding results, the output of many task requirements in autonomous driving can be completed, that is, the potential motion behavior of the vehicle can be predicted.

[0057] Among them, the above-mentioned autonomous driving model is applied to the GPU, which also solves the problem of data transmission between the CPU and the GPU during the autonomous driving process, thereby improving the data processing efficiency of the autonomous driving model and reducing the cost of data transmission.

[0058] Optionally, the autonomous driving model (or the end-to-end multi-task network after quantization) can be deployed on a designated computing chip (e.g., a designated on-chip computing unit on the vehicle side), and forward propagation is performed only on the GPU to complete the output of many task requirements in autonomous driving; therefore, in the embodiment of the present application, the architecture of the autonomous driving model can also be referred to as an end-to-end multi-task network model structure.

[0059] Exemplarily, the feature coding information includes but is not limited to: Bird's Eye View (BEV) feature map and voxel coding, wherein voxel coding is the coding of volumetric pixels; the decoding results include but are not limited to: three-dimensional (3D) targets, 3D motion and local vector fragments (i.e., local high-precision map vector fragments); the potential motion behaviors include but are not limited to: the 3D motion trajectory of the vehicle and the state control signal of the vehicle.

[0060] In an alternative implementation, see Figure 2 As shown, the autonomous driving model also includes: a global positioning module connected to the feature extraction and encoding module, and the global positioning module is used to determine the global positioning posture of the vehicle based on the posture information collected by multiple sensors of the vehicle.

[0061] Exemplarily, the above-mentioned multiple sensors include: an inertial measurement unit (IMU), a global navigation satellite system (GNSS) sensor and a wheel speed sensor; it should be noted that in the embodiments of the present application, no specific limitation is made on the category and specific type of the sensor.

[0062] Therefore, the global positioning posture of the vehicle comes from the high-precision positioning result of the vehicle (that is, the posture information collected by the above multiple sensors); see Figure 3 As shown in the figure, taking into account many low-cost sensor characteristics such as the integral error caused by IMU acceleration signal noise, the excellent real-time performance of IMU, the credibility of GNSS solutions in different scenarios, etc., a high-precision positioning tightly coupled framework is designed with IMU angular velocity and wheel speed integral as the predicted pose, GNSS, visual perception and high-precision map matching as the observed pose (optional, the algorithm can be switched freely), and the predicted pose and pose correction matrix jointly optimize the output.

[0063] Through the above embodiments, the posture information collected by multiple sensors of the vehicle is comprehensively considered, which improves the accuracy of the vehicle's global positioning posture and ensures the accuracy of the output results of many task requirements during the autonomous driving process to a certain extent.

[0064] In an optional implementation, Figure 2 As shown, the autonomous driving model also includes: a 3D vector reconstruction module, which is used to reconstruct a vector map around the vehicle's driving trajectory based on the global positioning posture and local vector fragments in the decoding results.

[0065] The local vector segment at least includes: the local positioning posture of the vehicle.

[0066] For example, see Figure 4 As shown in the figure, during the reconstruction of the 3D vector map, the matching sampling point pairs of the local vector fragments and the global positioning pose in the key frame can be obtained by timestamp interpolation, and an optimization graph can be constructed based on this. The optimization graph is iteratively optimized to generate a transformation matrix that can describe the coincidence of the local vector fragments to the global positioning pose, thereby fusing the information of the two and realizing real-time reconstruction of the vector map near the vehicle trajectory.

[0067] Through the above-mentioned embodiments, a vector map around the vehicle's driving trajectory is reconstructed according to the global positioning posture and the local vector fragments in the decoding results, thereby improving the accuracy of the real-time reconstructed vector map; and the vectorization results are output end-to-end, thereby reducing the data transmission cost.

[0068] In an alternative implementation, see Figure 5 As shown, the feature extraction and encoding module includes: an image preprocessing module and a temporal attention module; wherein the image preprocessing module is used to perform parameter calibration and feature extraction on the global positioning posture to obtain multiple posture feature maps, and the temporal attention module is used to perform temporal tracking on multiple posture feature maps.

[0069] Through the above embodiments, the preprocessing of the global positioning posture is achieved, and multiple more accurate posture feature maps are obtained. In addition, based on the original spatiotemporal sequence attention mechanism and flexible training inference network, the on-chip inference speed of the autonomous driving model is greatly improved without changing the accuracy.

[0070] In an optional implementation, the temporal attention module is further used to obtain feature encoding information according to a plurality of posture feature maps after temporal tracking, a set of feature sampling points associated with the vehicle type of the vehicle in the first preset index table, and a set of feature sampling point weights associated with the vehicle type in the second preset index table; illustratively, refer to Figure 6 and Figure 7 As shown in the figure, most of the use environments of autonomous driving vehicles are small displacement scenarios. Therefore, the core idea of ​​the time series attention module is to bring the characteristics of the historical frame into the current frame to complete the time series tracking. Considering the on-chip computing resources, the feature map is not transformed here, only the sampling reference point is transformed. In order to take into account the versatility of cloud training models and the efficiency of vehicle-side reasoning, for different models, the vehicle-side uses the method of querying the index table configured offline to replace complex calculations.

[0071] Through the above embodiment, according to the first preset index table and the second preset index table set for the vehicle type, the feature sampling point set and its corresponding feature sampling point weight set are queried, which not only improves the versatility of the autonomous driving model, but also greatly reduces the occupancy of on-chip computing resources and computational complexity, thereby improving the efficiency of vehicle-side reasoning.

[0072] In an optional implementation, the attention de-encoding module includes: multiple attention mechanism network layers, each attention mechanism network layer is trained according to a corresponding feature sampling point set and a preset loss function; it should be noted that in the embodiments of the present application, no specific limitation is imposed on the above-mentioned preset loss function.

[0073] For example, see Figure 8As shown in the figure, the attention de-coding module takes the attention mechanism under natural language as the core, and exchanges and fuses all image encoding information (i.e., feature encoding information) according to weights (i.e., feature sampling point weights) through a multi-layer / N-layer (e.g., 6-layer) attention mechanism, and outputs multi-task target results end-to-end; and, during training, each attention mechanism network layer in the attention de-coding module optimizes the back propagation process between multiple / N layers by combining hierarchical reference points (i.e., feature sampling point sets) and loss functions, so that the results do not have huge deviations, thereby ensuring the accuracy of the model.

[0074] Based on the above method, it is not difficult to know that the above weights are obtained through back-propagation training; it should be noted that the fused image encoding information is a multi-dimensional feature (e.g., 1000+ dimensions); it basically includes: color, gradient, edge, luminosity, grayscale, etc.

[0075] Optionally, the attention de-encoding module is also used to generate a 3D target and estimate the 3D motion trajectory of the 3D target, and determine the 3D occupied area of ​​the 3D target based on the 3D motion estimation result of the 3D target; wherein the 3D target is other road participants, such as pedestrians, non-motor vehicles, etc.

[0076] Through the above-mentioned embodiments, the generation of 3D motion trajectories, 3D motion estimation results and 3D occupied areas of other road participants is realized. In this way, during the automatic driving process, the status of other road participants is fully considered, further improving the accuracy of subsequent predictions of the vehicle's potential motion behavior.

[0077] In an optional implementation, Figure 2 As shown, the autonomous driving model also includes: a standard definition (SD) navigation map module and a vehicle dynamics constraint module connected to the graph neural network module; in this way, the autonomous driving model reduces its dependence on high-precision maps with the help of the SD navigation map and the vehicle dynamics constraint module in the SD navigation map module, and realizes the vehicle's autonomous driving based on "light maps", that is, through the graph neural network module, the potential movement behavior of the vehicle is accurately predicted.

[0078] For example, see Fig. 9 As shown in the figure, the graph neural network module builds associations between different input instances (such as SD navigation map line points, 3D occupied areas, and map vector fragments) through graph node information exchange; for example, after the gated recurrent encoder encodes the 3D occupied area and global positioning pose, and obtains the initial node, the probabilities of all potential future motion behaviors are obtained through the attention mechanism, and the optimal trajectory is selected for planning and output.

[0079] It should be noted that the above-mentioned association between instances through the exchange of graph node information is established through the global positioning posture.

[0080] Further, see Fig.10 As shown, it is a detailed design of a pure spatiotemporal end-to-end autonomous driving model provided in an embodiment of the present application. The autonomous driving model can realize local high-precision map vector fragment detection, 3D detection and trajectory prediction of other road participants (corresponding to 3D target generation, 3D motion estimation and 3D occupied area), 3D vector map reconstruction (corresponding to 3D vector reconstruction module), 3D space path planning (i.e., 3D motion trajectory generation of the vehicle) and vehicle status control signal prediction, etc. based on continuous surround video frames, positioning posture and SD navigation map.

[0081] See also Fig.11 As shown, the embodiment of the present application also provides an autonomous driving method of an autonomous driving model, which is applied to a GPU. The execution subject can be an autonomous driving model, a server deploying an autonomous driving model, or other network devices. In the embodiment of the present application, taking the server deploying the autonomous driving model as an example, the specific process of the method is as follows:

[0082] S1101: In response to the autonomous driving instruction of the target object, obtain driving information of the vehicle associated with the autonomous driving instruction.

[0083] Exemplarily, the driving information of the above-mentioned vehicle includes, but is not limited to: multiple positioning posture information of the vehicle (posture information collected by multiple sensors of the vehicle respectively) and environmental information around the vehicle.

[0084] S1102: Input the vehicle's driving information into a pre-trained autonomous driving model to obtain the vehicle's autonomous driving strategy.

[0085] Among them, the autonomous driving strategy includes: the vehicle's potential movement behavior.

[0086] In an optional implementation, when executing step S1102, after obtaining the driving information of the vehicle, the server can parse the driving information through a pre-trained autonomous driving model to obtain multiple posture information of the vehicle at the current moment and the environmental information of the vehicle at the current moment, thereby determining the global positioning posture of the vehicle based on the multiple posture information, and generating the vehicle's autonomous driving strategy based on the global positioning posture and environmental information to achieve safe and reliable autonomous driving of the vehicle.

[0087] S1103: Based on the autonomous driving strategy, instruct the vehicle to perform autonomous driving.

[0088] In an optional implementation, during the execution of step S1103, the server receives the target object's instruction to display the driving conditions of the vehicle, and can respond to the target object's instruction to display the driving conditions of the vehicle, instructing the vehicle to display the 3D motion trajectory and status control signals within a set time range in the future (for example, the next 20 seconds, 1 minute, etc.); in this way, not only is the visualization of the vehicle's expected operating trajectory and status control signals in the future period of time based on the autonomous driving strategy achieved, but the experience of the people on the vehicle is also improved to a certain extent.

[0089] To summarize, the autonomous driving method based on the autonomous driving model recorded in the above steps S1101 to S1103 responds to the autonomous driving instructions of the target object, obtains the driving information of the vehicle associated with the autonomous driving instructions, and then inputs the driving information of the vehicle into the pre-trained autonomous driving model, obtains the autonomous driving strategy including the potential motion behavior of the vehicle, and then instructs the vehicle to perform autonomous driving based on the obtained autonomous driving strategy; in this way, only forward propagation on the GPU is required to complete the output of many task requirements in autonomous driving, and the problem of data transmission between the CPU and GPU during the autonomous driving process is solved, thereby improving the data processing efficiency of the autonomous driving model and reducing the cost of data transmission.

[0090] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. An autonomous driving model, It is characterized in that Applied to GPU, including feature extraction and encoding module, attention de-encoding module and graph neural network module; The feature extraction and encoding module is connected to the attention de-encoding module; wherein the feature extraction and encoding module is used to extract and encode the driving information of the vehicle, and the attention de-encoding is used to decode the feature encoding information output by the feature extraction and encoding module; The attention de-coding module is connected to the graph neural network module; wherein, the graph neural network module is used to predict the potential movement behavior of the vehicle based on the decoding result output by the attention de-coding module.

2. The autonomous driving model according to claim 1, It is characterized in that The autonomous driving model also includes: a global positioning module connected to the feature extraction and encoding module, and the global positioning module is used to determine the global positioning posture of the vehicle based on the posture information collected by multiple sensors of the vehicle.

3. The autonomous driving model according to claim 2, It is characterized in that The autonomous driving model also includes: a three-dimensional 3D vector reconstruction module, which is used to reconstruct a vector map around the vehicle's driving trajectory based on the global positioning posture and the local vector fragments in the decoding results; wherein the local vector fragments at least include: the local positioning posture of the vehicle.

4. The autonomous driving model according to claim 2, It is characterized in that The feature extraction and encoding module includes: an image preprocessing module and a temporal attention module; wherein the image preprocessing module is used to perform parameter calibration and feature extraction on the global positioning posture to obtain multiple posture feature maps, and the temporal attention module is used to perform temporal tracking on the multiple posture feature maps.

5. The automatic driving model as claimed in claim 4, It is characterized in that The temporal attention module is also used to obtain the feature coding information based on multiple posture feature maps after temporal tracking, a set of feature sampling points associated with the vehicle type of the vehicle in a first preset index table, and a set of feature sampling point weights associated with the vehicle type in a second preset index table.

6. The autonomous driving model according to claim 1, It is characterized in that The attention de-encoding module includes multiple attention mechanism network layers, each of which is trained based on a corresponding feature sampling point set and a preset loss function.

7. The automatic driving model according to claim 1, It is characterized in that The attention de-encoding module is also used to generate a 3D target and estimate the 3D motion trajectory of the 3D target, and determine the 3D occupied area of ​​the 3D target based on the 3D motion estimation result of the 3D target.

8. An automatic driving method of the automatic driving model according to any one of claims 1 to 7, It is characterized in that Applied to GPU, including: In response to an autonomous driving instruction of a target object, obtaining driving information of a vehicle associated with the autonomous driving instruction; Inputting the driving information of the vehicle into a pre-trained autonomous driving model to obtain an autonomous driving strategy for the vehicle; wherein the autonomous driving strategy includes: potential motion behaviors of the vehicle; Based on the autonomous driving strategy, instruct the vehicle to perform autonomous driving.

9. The method according to claim 8, It is characterized in that The step of inputting the driving information of the vehicle into a pre-trained automatic driving model to obtain the automatic driving strategy of the vehicle includes: Parsing the driving information to obtain a plurality of position information of the vehicle at the current moment and information about the environment in which the vehicle is located at the current moment; Based on the multiple pose information, the global positioning pose of the vehicle is determined, and based on the global positioning pose and the environmental information, an automatic driving strategy for the vehicle is generated.

10. The method according to claim 8, It is characterized in that The process of instructing the vehicle to perform automatic driving based on the automatic driving strategy further includes: In response to the target object's instruction to display the driving road condition of the vehicle, the vehicle is instructed to display a 3D motion trajectory and a state control signal within a future set time range.

Images